FM04 12:lO Proceedings of the 37th IEEE Conference on Decision & Control Tampa, Florida USA December 1998 Optimal Robust Disturbance Attenuation for Continuous Time-Varying Systems Mohamed Seddik Djouadi Centre for Intelligent Machines, and Dept. of Electrical Eng., McGill University 3480 University Street Montrkal, Qukbec, Canada Corresp. address Dept. of Electrical Eng. Wayne State University 5050 Anthony Wayne Drive # 3100 Detroit, Michigan, USA 48202 djouadi@ece.eng.wayne.edu, djouadi@cim.mcgill.ca and for h < 0 by: Abstract Dh : In this paper we consider the Optimal Robust Disturbance Attenuation Problem (ORDAP) for continuous time-varying systems subject to timevarying unstructured uncertainty. We show that for causal (possibly time-varying) continuous systems, ORDAP is equivalent to finding the smallest fixed point of a “two-disk” type optimization problem under time-varying feedback control laws. Duality is applied in the context of nest algebra of causal stable systems, to prove existence of optimal continuous time-varying controllers. We also show that for time-invariant nominal plants, time-varying control laws offer no improvement over time-invariant feedback control laws and hence generalizing previous results obtained for discrete time-varying systems Definitions and Notation B ( E , F ) denote the space of bounded linear operators from a Banach space E to a Banach space F , endowed with the operator norm. L”0, CO) the standard Lebesgue space of essentially square integrable functions defined on the interval [O,m). L2[0,00) --+ L2[[o,00) f(t) --+f ( t - h ) 0-7803-4394-8198 $10.000 1998 IEEE (1) (2) At) -4 f(t - h)X[O,m)( t ) where X [ O , ~ is ) the characteristic function of the interval [0,00).An operator A E ’B(E,F ) is said to be time-invariant if, for all h 2 0, A satisfies the operator equation [l] P, the usual truncation operator, which sets all outputs after time T to zero. 0 An operator A E % ( E ,F ) is said to be causal if it satisfies the operator equation: P,AP, = PTA, Vr E (0,m) (4) The subscripts “sc”, ‘‘c” and the superscript ‘‘ti” denote the restriction of a subspace of operators to its intersection with causal, strictly causal (see [2] for definition), and time-invariant operators respectively. * stands for the adjoint of an operator or the dual space of a Banach space depending on the context. -* denotes convergence in the weak* topology. Dh the family of delay operators defined for h 2 0 by: Dh : L 2 [ 0 , ~-+ ) L2[o,m) 1 Introduction The Optimal Robust Disturbance Attenuation Problem (ORDAP) was formulated by Zames 131, and considered by Bird, Francis, Owen and Zames 381 9 If a particular controller Q achieves a "worst-case" weighted sensitivity function less than some r > 0 then, [4, 5, 6, 7, 81. In ORDAP a stable uncertain plant P is subject to disturbances at the output. The objective is t o find a feedback control law which provides the best uniform attenuation of uncertain output disturbances in spite of uncertainty in the plant model. Here we examine the ORDAP for possibly time-varying continuous systems and feedback control laws. Analysis of time-varying control strategies for optimal disturbance rejection for known time-invariant plants has been studied by Shamma and M.A. Dahleh [9], Chapellat and M. Dahleh [l]. A robust version of these problems were considered in [lo, 11, 121 in different induced norm topologies. They showed that for time-invariant nominal plants, time-varying control laws offer no advantage over time-invariant ones. The ORDAP for discrete timevarying systems and feedback laws has been studied by Owen [7,13]. He characterized the problem in a predual space and proved that in some induced norm topology the same conclusion as before holds. In this paper we analyse the ORDAP for continuous time-varying systems extending previous results and therefore settling an open question in [7]. I\W(I - P o Q ) ( I + XVPoQ)-'I1 (I XVP,Q)-l E B c ( L 2 ,L'), vx E IB,(L"LL'), llxll < 1 I r , and + (8) Expression (8) is equivalent to: Ilw(I- P o Q ) ( I + xVf'oQ)-'fllp Lrllfll~z, and ( I f XVP,Q)-l E 'Bc(L2,LL2), (9) vx E B,(L"L'), llXll < 1, Vf E L"0,oo) yielding I l W ( I - P 0 Q ) F l l p I rll(I + X V P o Q ) F I I ~ 2 and , ( I XVPoQ)-l E B,(L2, L'), VX E !Bc(L2,L'), + IlXll < 1, V F E L"O,oo), IIFllU L 1 (10) which is in turn implied by (see chapter 5 of [ 2 ] ) : Ilw(1-PoQ)FlILz I r l l F l l ~ 2- r l l v p o Q F I I ~ ~ , V F E L 2 , llFll~z5 1 (11) Hence: 2 Problem Formulation Let Po E IBB,,(L2[0, m),L2[0,m))be the nominal (possibly time-varying) plant, and denote the set of plant uncertainty by: e ( P o , V ) = { P E IBc(L"-oo,m),L"-oo,oo)) : P = XVP, Po, (5) E IB,(L"-m, m),L"-oo, m)),J J X< JJ 1) + x where V is a causal, linear time-invariant weighting function. The ORDAP can be shown t o be equivalent to finding the optimal worst case sensitivity function with respect to disturbances and plants in e(P0,V ) , achievable by a feedback control law. Mathematically this is equivalent to Po = inf sup (1 IIW(I+ C stabilizing PEC(PO,V) p E e(po,v) inf Q E %(L2, L2) ( I + XVP,Q)-' E IB(L"LL") ((W(I PoQ)(I + XVPoQ)-lII + r IIVPoQfllLa ) then [ is a continuous, positive, non-decreasing function of T . We get a similar theorem to theorem 6.1 for discrete time-varying systems in [7]: Theorem 1 1) Let Po be time-invariant and po as above, if there exists an optimal Q E 'Bc(L2,L 2 ) for each r E [0,11 in the expression (141, then po is equal to the smallest f i e d point of [ ( r ). 2) If Po is time-varying then po is bounded above b y the smallest fixed point of [ ( r ) . Proof: similar t o the discrete time case (see [14]). Hence the optimization (7) for time-invariant nominal plant reduces to: (6) where W is a causal, linear, time-invariant weighting function. Expression (6) can be written as: Po = (IlW(I - P o Q ) f l l u SUP PE C(P0 , V ) (7) 3820 Next we proceed t o give the duality structure of the problem, which shows existence of an optimal Q for P O . Let lL2 be the Banach space L2 x L2 under the norm: 3 Duality Structure and Existence of an The following lemma characterizing the dual space of C2 follows easily from [li']: Optimal Solution Let C2 be the Banach space L2[0,CO) x L2[0,00) under the norm: The vector function ( Lemma 1 ) a (L2)* N C2 a (C2)* N IL2 w ( ~ ~ o can ~ Qbe ) Hence all these Banach spaces are reflexive. viewed as an operator fr'om L2[0,co)into C2 with the operator induced norm: Define N(B1,B2) t o be the Banach space of nuclear operators mapping the Banach space B1 to the Banach space B2 under the nuclear norm. Recall that an operator A : B1 -+ B2 is said to be nuclear if it has the representation [MI: Therefore, po can be expressed as a distance prob- ( y ) IB(L2,P) = ( 7 ) Po B c ( L 2 , L 2 ) of lem from the vector function to the subspace S n + where < > is the duality product and the nuclear norm is defined to be: a , . + ( ) Al'i71, IIAtInuc E IICII. ItenIII (21) where the infimum is taken over all possible representations of A. The trace of the nuclear operator A is denoted by tr(A) and is defined by assumption !Bc(L2,L 2 ) =inf{C n (A) implies that the operator R*R E B ( L 2 , L 2 ) has a bounded inverse, this ensures closedness of S According t o Arveson ([16], Corollary 2, see also [2]), the self-adjoint operator R* R has a spectral factorization of the form: R"R = A"& where A , A-' (20) n B ( L 2 , L 2 )We ) . assume that IW(jw)12 IV(jw)12> 0 , 0 5 w 5 00. Then there exists an outer spectral factor A1 E H m , invertible in Hm, such that lA1(jw)l2 = IW(jw)12 lV(jw)12. Therefore Alpo is a bounded linear operator in B c ( L 2 L , 2 )has a polar decomposition UlG, where U1 is a partial isometry and G a positive operator both defined on L2 [15]. Next we assume (A) U1 is unitary and G f l E IB(L2,L2). Let R = llF,*Il . IlenII < 00 F,* E B,*,and of < F,*,Fn > tr(A) = n This sum is well defined and can be shown to be independent of the representation [ 181. The following lemma applies to the Banach spaces L2 and lL2. Define Lemma 2 [18, 11 R2 = RA-' then R;R2 = I , and S has the equivalent representation, S = R2Bc(L2,L2)). After absorbing A in the free parameter Q, the optimization problem (14) is then equivalent to: B(L2, C2) Y N ( P ,P)* i f 4 E N ( L 2 , L 2 ) * the , isometric isomorphism given by: $(A) =< B , A >= trA*B = trB*A, where A E N ( L 2 , L 2 ) , B E B ( L 2 , f ? ) 3821 (23) 2~ is Next, we characterize the preannihilator of the the Banach space of causal operators 'B,(L2,L 2 ) as a subspace of B ( L 2 , L 2 )in the space of nuclear operators N ( L 2 , L 2 ) . But first define for each t E (-00, CO), Pt ( L 2 )the subspace of L2 consisting of all functions f in L2 such that f = 0, a.e. on ( t , 00). Then A4 = { P t ( L z ) , -00 < t < 0 0 ) is a continuous nest [15]. The space !J3,(L2,L 2 ) may then be viewed as a nest algebra since causal operators leave M invariant, i.e., AMt C Mt, for all Mt E M . The Banach space N ( L 2 , L 2 ) is nothing but the well-known space of trace-class operators from L2 t o L 2 , and that the nuclear norm in this case reduces t o the trace-norm [9]. Applying theorem 16.6 [15], we get the following lemma. Lemma 3 N c ( L 2 , L 2 ) is the preannihilator of 23@, L"). Proof the preannihilator of Using a standard result from Banach space duality theory relating the distance from a vector to a subspace and an extremal functional in the predual (see [19]), we deduce the following theorem. Theorem 2 Under assumption ( A ) there exists at least one Q o E 'B,(L',L2), i.e., a linear timevarying control law such that: A E IS 23,(L2,L'), 4 Time-Varying versus Time-Invariant ' B , ( L 2 , L 2 ) is defined by Control laws ''Bc(L2,L 2 ) = {A E N ( L 2 , L 2 ) : t r ( T A ) = 0 , VT E 'B,(L2,L')} (24) Let M; be the closed linear span of { M , E M : 7 < t } , but since M is a continuous nest, then M; = Mt. It follows by theorem 16.6 [15] that ' S , ( L 2 ,L 2 ) = N c ( L 2 , L'). Now let S, be the subspace of N(C2,L2(0,cm))) defined as follows: S, = { p ( Q ( I - R2R;) + AR;) (La : @ E N(,CE,L'), A In the next theorem, we show that no advantage is gained in performance if we allow our controllers to be time-varying, and thus generalizing the same result obtained for discrete time systems in [7, 131. (25) E N c ( L 2 ,L')} where p is the canonical projection of L2(--oo,00) into L 2 [ 0 , m ) ,and CL is defined t o be the Banach space L 2 ( - m , 00) x L ' ( - ~ o ,00) under the norm of ,!?. Define the following subspace of N ( L 2 , L 2 ) : IS = { A E N(L2,1L2): A* E S,} Lemma 4 IS is the preannihilator of S In this section, we assume that the nominal plant Po is time-invariant. It follows also that Rz is also time-invariant. Define the following performance index when the controllers are constrained to be linear time-invariant: Theorem 3 If Po k linear, causal, time invariant and assumption ( A ) holds, then P o = P: (26) (29) Proof since 23p (L" L 2 ) c 'Bc(L" L 2 ) ,it is obvious that p o 5 p z . Let Q E 'Bc(L2,L2) in N (L 2 ,L2). Proof let T E 'B(L2(O,0o),f?), then <T,A>=O, V A E ' S ~ such a Q exist by theorem 2. Using an idea of Chapellat and M. Dahleh [l],let h > 0. Since llD,(( 5 1, Vu, we have that Vn > 0 + t r ( p ( @ ( I - RzRI;)(,,T pNR;I,z 7') = 0 V* E N ( C 5 , L 2 [ O , o o ) ) , NE N c ( L 2 , L 2 ) U ( I - R2R;)IpTp = 0 and R 2 R ; l p T p E B C ( L 2L, 2 ) by lemma 3 3822 For all h Using time-invariance of W and R2, we get Define for all n 2 0 . n then (32) yields IlQn(h)Il 5 11Q11, vn 2 0 and then llR2Qn(h)ll is uniformly bounded by 11R21111Q11. Moreover the predual space N(L2,1L2)of B ( L 2 , L 2 ) ,by Alaoglu’s theorem there exist a subsequence of {R2Qnk( h ) } which converges in the weak* topology to some vector function R E B ( L 2 ,C 2 ) ,we write by (34) and a property of weak* limits, we have By definition of the weak* limit, V A E’ S , 0 =< R 2 Q n k ( h ) , A>+< R , A >, as k + 00, hence R = R z Q ( h ) , @ ( h ) E IB,(L2,LL”).The adjoint of Dh, D i is defined by Vf E L2 ( D f i f ) ( t= ) f ( t-k h ) = ( D - h f ) ( t ) Now V A E N (L” lL2 ), we have (37) > 0, we showed that Qn 2 1, define the sequence R2Qn = R2Q(:). By-(39) it is bounded, therefore there 3Q s.t. R2Qnk -* R2Q and R2Q satifies (40) as before. Again using properties of nuclear operators it is easy to see that R2QnkDh - DhR2Qnk -* R2QDh - R2Q. Using exactly the same procedure as in [I] to prove that R2QnkDh - DhR2Qnk -* 0 shows that R2Q is time-invariant, but since R2 is time-invariant, Q must also be time-invariant and this completes the proof. References [l] H. Chapellat and M. Dahleh. Analysis of time-varying control strategies for optimal disturbance rejection and robustness. IEEE Transactions on Automatic Control, 37( 11):1734-1746, 1992. [a] A. Feintuch and R. Saeks. System Theory: A Hilbert Space Approach. Academic Press, New York, London, 1982. [3] G. Zames. Feedback and optimal sensitivity: Model reference transformations, multiplicative seminorms, and approximate inverses. IEEE Transactions on Automatic Control, AC26(2):301-320, April 1981. [4] J.F Bird and B.A. Francis. On the robust disturbance attenuation. Proceeding of IEEE Conference on Decision and Control, pages 1804-1809, 1986. [5] B.A. Francis. On disturbance attenuation with plant uncertainty. Workshop on New Perspectives in Industrial Control System Design, 1986. [6] G. Zames and J.G. Owen. Duality theory for MIMO robust disturbance rejection. IEEE Transactions on Automatic Control, AC-38(5):743-752, May 1993. [7] J.G. Owen. Performance Optimization of Highly Uncertain Systems in Hm. PhD thesis, Dept. of Elect,rical Eng., McGill University, 1993. [8] J.G. Owen and G. Zames. Robust disturbance minimization by duality. Systems and Control Letters, 19:255--263, 1992. [9] J.S. Shamma and M A . Dahleh. Timevarying versus time-invariant compensation for rejection of persistent bounded disturbances and robust stabilization. IEEE Transactions on Automatic Control, 36(7):838-847, 1991. 3823 110) J.S. Shammz. Robust stability with timevarying structured uncertainty. Proceedings of the CDC, Tucson, Arizona, 1992. [ll] M. Khammash and M. Dahleh. Time-varying control and the robust performance of systems with structured norm-bounded perturbations. Proceedings of the CDC, Brighton, England, 1991. [12] M. Khammash and J.B. Pearson. Performance robustness of discrete-time systems with structured uncertainty. IEEE Transactions on Automatic Control, 36(4):398412, 1991. [13] J.G. Owen. Robust performance for timevarying systems: A functional analytic approach. Proceedings of the ACC, San Fransisco, California, 19, 1993. [14] M.S. Djouadi. Optimization of Highly Uncertain Feedback Systems in Hm. P h D thesis, Dept. of Electrical Eng., McGill University, 1998. [15] K.R. Davidson. Nest algebras. Longman Scientific & Technical, UK, 1988. [16] W. Arveson. Interpolation problems in nest algebras. Journal of Functional Analysis, 4:67-71, 1975. [17] J. DieudonnBe. Sur le theorbme de Lebesgue Nikodym V. Canadian Journal of Mathematics, 3:129-139, 1951. [18] J . Diestel and J.J. Uhl. Vector Measures. AMs, Providence, RI, 1977. I191 D.G. Luenberger. Optimization by Vector Space Methods. John Wiley, 1968. 3824